Difference between revisions of "Config:InitialDesign"

From SUMOwiki
Jump to navigationJump to search
 
(2 intermediate revisions by the same user not shown)
Line 1: Line 1:
 +
The InitialDesign components lets you specify the starting data for the sequential building process or if you are using a one-shot approach to specify the training data. You can either specify a sampling design, i.e. the locations that need to be sampled and evaluated or supply data with the output values. The SUMO Toolbox implements several sampling strategies which are listed below.
 +
 +
 
'''Generated for SUMO toolbox version 7.0'''.
 
'''Generated for SUMO toolbox version 7.0'''.
 
''We are well aware that documentation is not always complete and possibly even out of date in some cases. We try to document everything as best we can but much is limited by available time and manpower.  We are are a university research group after all. The most up to date documentation can always be found (if not here) in the default.xml configuration file and, of course, in the source files.  If something is unclear please dont hesitate to [[Reporting problems|ask]].''
 
''We are well aware that documentation is not always complete and possibly even out of date in some cases. We try to document everything as best we can but much is limited by available time and manpower.  We are are a university research group after all. The most up to date documentation can always be found (if not here) in the default.xml configuration file and, of course, in the source files.  If something is unclear please dont hesitate to [[Reporting problems|ask]].''
Line 4: Line 7:
  
 
=== lhd ===
 
=== lhd ===
Latin Hypercube desing of experiments (DOE). Choose an initial sample set in such a way that they form a latin hypercube. This initial design will first try to download a pre-optimized maximin Latin hypercube from the website [http://www.spacefillingdesigns.nl]. If such a design is not available, or no proper connection to the website can be made, the Latin hypercube is generated by the toolbox. Note that this Latin hypercube may be suboptimal, and may leave considerable gaps in the design space.
+
Latin Hypercube desing of experiments (DOE). Chooses an initial sample set in such a way that they form a latin hypercube. This initial design will first try to download a pre-optimized maximin Latin hypercube from the website [http://www.spacefillingdesigns.nl]. If such a design is not available, or no proper connection to the website can be made, the Latin hypercube is generated by the toolbox. Note that this Latin hypercube may be suboptimal, and may leave considerable gaps in the design space.
  
 
<source xmlns:saxon="http://icl.com/saxon" lang="xml">
 
<source xmlns:saxon="http://icl.com/saxon" lang="xml">
Line 33: Line 36:
 
</[[Config:InitialDesign|InitialDesign]]>
 
</[[Config:InitialDesign|InitialDesign]]>
 
</source>
 
</source>
=== lhdWithCornerPoints ===
+
 
 +
=== dataset ===
 +
Reads an initial design from a dataset.  Each row should contain one sample. The inputvalues come first, the outputvalues, if there are any, come second.
 +
 
 +
If only input values are supplied (hasOutputs=false), each value has to be in the fixed [-1,1] range (model space). The samples are later automatically transformed to their original range before they are passed to the sample evaluator. This is done to easily support different ranges for variables without having to change the initial dataset design.
 +
 
 +
If, however, output values are also provided (hasOutputs=true), the samples are instantly fed to the toolbox for modelling, since they don't have to be evaluated anymore. Because of this property, the inputs have to be in the original range defined in the simulator xml file, instead of the fixed [-1,1] range.
 +
 
 +
<source xmlns:saxon="http://icl.com/saxon" lang="xml">
 +
<[[Config:InitialDesign|InitialDesign]] type="[[InitialDesign#DatasetDesign|DatasetDesign]]">
 +
  <!-- Where should we load the dataset from? -->
 +
  <Option key="filename" value="/path/to/your/dataset.txt"/>
 +
  <!-- Does the dataset also contain responses? (instead of only the inputs)
 +
        If so, hasOutputs must be set to yes and the range of the inputs in the file must
 +
        match the range of the simulator.  If not, the range of the inputs must be [-1 1].
 +
  -->
 +
  <Option key="hasOutputs" value="no"/>
 +
</[[Config:InitialDesign|InitialDesign]]>
 +
</source>
 +
 
 +
=== CombinedDesign ===
 +
 
 +
The CombinedDesign lets you combine different initial designs. For example, here is a configuration where a normal Latin HyperCube design is combined with a 2-level Factorial Design, which adds the corner points of the input space to the original Latin HyperCube.
 +
 
 
Specifies a combined Latin HyperCube and FactorialDesign
 
Specifies a combined Latin HyperCube and FactorialDesign
 
<source xmlns:saxon="http://icl.com/saxon" lang="xml">
 
<source xmlns:saxon="http://icl.com/saxon" lang="xml">
Line 51: Line 77:
 
       <Option key="levels" value="2"/>
 
       <Option key="levels" value="2"/>
 
   </[[Config:InitialDesign|InitialDesign]]>
 
   </[[Config:InitialDesign|InitialDesign]]>
</[[Config:InitialDesign|InitialDesign]]>
 
</source>
 
=== dataset ===
 
Reads an initial design from a dataset.  Each row should contain one sample. The inputvalues come first, the outputvalues, if there are any, come second.
 
 
If only input values are supplied (hasOutputs=false), each value has to be in the fixed [-1,1] range (model space). The samples are later automatically transformed to their original range before they are passed to the sample evaluator. This is done to easily support different ranges for variables without having to change the initial dataset design.
 
 
If, however, output values are also provided (hasOutputs=true), the samples are instantly fed to the toolbox for modelling, since they don't have to be evaluated anymore. Because of this property, the inputs have to be in the original range defined in the simulator xml file, instead of the fixed [-1,1] range.
 
 
<source xmlns:saxon="http://icl.com/saxon" lang="xml">
 
<[[Config:InitialDesign|InitialDesign]] type="[[InitialDesign#DatasetDesign|DatasetDesign]]">
 
  <!-- Where should we load the dataset from? -->
 
  <Option key="filename" value="/path/to/your/dataset.txt"/>
 
  <!-- Does the dataset also contain responses? (instead of only the inputs)
 
        If so, hasOutputs must be set to yes and the range of the inputs in the file must
 
        match the range of the simulator.  If not, the range of the inputs must be [-1 1].
 
  -->
 
  <Option key="hasOutputs" value="no"/>
 
 
</[[Config:InitialDesign|InitialDesign]]>
 
</[[Config:InitialDesign|InitialDesign]]>
 
</source>
 
</source>

Latest revision as of 14:26, 30 January 2012

The InitialDesign components lets you specify the starting data for the sequential building process or if you are using a one-shot approach to specify the training data. You can either specify a sampling design, i.e. the locations that need to be sampled and evaluated or supply data with the output values. The SUMO Toolbox implements several sampling strategies which are listed below.


Generated for SUMO toolbox version 7.0. We are well aware that documentation is not always complete and possibly even out of date in some cases. We try to document everything as best we can but much is limited by available time and manpower. We are are a university research group after all. The most up to date documentation can always be found (if not here) in the default.xml configuration file and, of course, in the source files. If something is unclear please dont hesitate to ask.

InitialDesign

lhd

Latin Hypercube desing of experiments (DOE). Chooses an initial sample set in such a way that they form a latin hypercube. This initial design will first try to download a pre-optimized maximin Latin hypercube from the website [1]. If such a design is not available, or no proper connection to the website can be made, the Latin hypercube is generated by the toolbox. Note that this Latin hypercube may be suboptimal, and may leave considerable gaps in the design space.

<[[Config:InitialDesign|InitialDesign]] type="[[InitialDesign#LatinHypercubeDesign|LatinHypercubeDesign]]">
   <!-- how many points to generate -->
   <Option key="points" value="20"/>
   <!--<Option key="weight" value="0.5"/>-->
   <!--<Option key="coolingFactor" value="0.9"/>-->
   <!--<Option key="p" value="5.0"/>-->
</[[Config:InitialDesign|InitialDesign]]>

factorial

Specifies a simple Factorial Design (uniform grid). The levels option determines the number of points per dimension. If only one value is given this value is used for every input dimension.

<[[Config:InitialDesign|InitialDesign]] type="[[InitialDesign#FactorialDesign|FactorialDesign]]">
   <!-- how many points to generate for each dimension as a vector -->
   <!-- a scalar value (l) is the same as [l l ... l] (length of input dimension) -->
   <Option key="levels" value="3"/>
</[[Config:InitialDesign|InitialDesign]]>

random

Specifies a trivial Random design

<[[Config:InitialDesign|InitialDesign]] type="[[InitialDesign#RandomDesign|RandomDesign]]">
   <Option key="points" value="20"/>
</[[Config:InitialDesign|InitialDesign]]>

dataset

Reads an initial design from a dataset. Each row should contain one sample. The inputvalues come first, the outputvalues, if there are any, come second.

If only input values are supplied (hasOutputs=false), each value has to be in the fixed [-1,1] range (model space). The samples are later automatically transformed to their original range before they are passed to the sample evaluator. This is done to easily support different ranges for variables without having to change the initial dataset design.

If, however, output values are also provided (hasOutputs=true), the samples are instantly fed to the toolbox for modelling, since they don't have to be evaluated anymore. Because of this property, the inputs have to be in the original range defined in the simulator xml file, instead of the fixed [-1,1] range.

<[[Config:InitialDesign|InitialDesign]] type="[[InitialDesign#DatasetDesign|DatasetDesign]]">
   <!-- Where should we load the dataset from? -->
   <Option key="filename" value="/path/to/your/dataset.txt"/>
   <!-- Does the dataset also contain responses? (instead of only the inputs) 
        If so, hasOutputs must be set to yes and the range of the inputs in the file must
        match the range of the simulator.  If not, the range of the inputs must be [-1 1].
   -->
   <Option key="hasOutputs" value="no"/>
</[[Config:InitialDesign|InitialDesign]]>

CombinedDesign

The CombinedDesign lets you combine different initial designs. For example, here is a configuration where a normal Latin HyperCube design is combined with a 2-level Factorial Design, which adds the corner points of the input space to the original Latin HyperCube.

Specifies a combined Latin HyperCube and FactorialDesign

<[[Config:InitialDesign|InitialDesign]] type="[[InitialDesign#CombinedDesign|CombinedDesign]]">
   <!-- Select samples in a Latin Hypercube Design -->
   <[[Config:InitialDesign|InitialDesign]] type="[[InitialDesign#LatinHypercubeDesign|LatinHypercubeDesign]]">
      <!-- how many points to generate -->
      <Option key="points" value="20"/>
      <!--<Option key="weight" value="0.5"/>-->
      <!--<Option key="coolingFactor" value="0.9"/>-->
      <!--<Option key="p" value="5.0"/>-->
   </[[Config:InitialDesign|InitialDesign]]>

   <[[Config:InitialDesign|InitialDesign]] type="[[InitialDesign#FactorialDesign|FactorialDesign]]">
      <!-- how many points to generate for each dimension as a vector -->
      <!-- a scalar value (l) is the same as [l l ... l] (length of input dimension) -->
      <Option key="levels" value="2"/>
   </[[Config:InitialDesign|InitialDesign]]>
</[[Config:InitialDesign|InitialDesign]]>